Policy
When the Tool You Banned Is Someone's Assistive Technology
When a candidate says AI is how they manage a disability and your round bans it, treat the sentence as a request for reasonable accommodation and start with one question: what is this round meant to measure? A model that drafts, reads aloud or takes dictation leaves a reasoning exercise intact, so allow it. A model doing the reasoning the exercise exists to watch does not, so offer another route to the same measurement before you deny anything.
The takeThe instinct is to protect comparability, and it points at the wrong object. Comparability belongs to the thing the round measures, never to the equipment on the desk. A round that bans a general-purpose model for everyone and then grades domain judgment has already broken its own comparability, because it is quietly scoring reading and typing alongside the thing it cares about. The request is doing you a favour. It is the first time anyone has asked what the round is for, and most rounds cannot answer.
Where Olive fits
Open a role and see what the work shows
Olive produces no composite, and no decision about a candidate is made by a machine alone: a person writes or edits each of the six findings, every finding carries the timestamped excerpt it rests on, and a released report exports with its rubric, scorer and bank versions attached. The session is asynchronous and runs on the candidate's own clock, so the candidate chooses when to take it; a change to its format or length is still a request to handle in the interactive process.
Assess your shortlistWhat is the round measuring, and does the model do that?
Whatever the round measures is the only thing the ban gets to protect. The ADA's test-administration rule, 29 CFR 1630.11, ends on a parenthetical that frames this case. Where a test is given to an applicant whose disability impairs sensory, manual or speaking skills, the results have to reflect what the test purports to measure rather than the impairment. The exception is where those skills are the factors the test purports to measure 1.
Read the exception rather than the rule. The exception says the equipment question stays open until the purpose question closes.
The rule is narrower than the principle it carries. 29 CFR 1630.11 governs only how tests are administered, and only where a disability impairs sensory, manual or speaking skills. The duty to accommodate and the ban on screening people out sit elsewhere in the statute.
The screen-out ban is where a round like yours actually lands. Enacted with the ADA in July 1990, 42 U.S.C. 12112(b)(6) makes it discrimination to use a qualification standard, employment test or other selection criterion that screens out or tends to screen out an individual with a disability. The ban lifts only where the criterion, as the employer uses it, is job-related for the position in question and consistent with business necessity 2.
The word doing the work is individual. The provision reaches one person, so a round with a healthy overall pass rate can still be a problem for the candidate in front of you.
Which supports leave the measurement alone?
Most of them, on most rounds. Dictation, text-to-speech, a spelling pass, a model that turns a wall of instructions into a checklist: none of those performs the domain judgment a reasoning exercise exists to watch. Permitting them moves nothing you were trying to score.
Extended time and an alternative version of a test, including one compatible with a screen reader, are named in the EEOC's 2022 technical assistance among accommodations that may be effective 3.
The document is agency technical assistance rather than law, and it reads here through an Internet Archive capture. The EEOC took both of its AI technical assistance documents off eeoc.gov in late January 2025, and neither has come back 4.
One kind of request does touch the measurement. A candidate asking to run the reasoning through a model, on an exercise built to watch them reason, is asking for the measured act to be done by something else.
The hard case is narrower than the worry about it. Where a round banned the tool without ever writing down what the ban protects, the first honest answer is often that the model does not perform the measured act at all.
Three requests, three answers, on an exercise that measures how a candidate frames a problem and checks a claim:
- Read the brief aloud and dictate the answer. The exercise never claimed to measure reading or typing. Allow it, and put the permission in writing.
- Draft a first version with a model, then edit it. Drafting is neither framing nor checking. Allow it, and grade the edit.
- Ask the model which of two approaches is better and follow the answer. Framing is the measured act here. Deny that one, and offer a format that reaches framing another way.
The three answers differ because the exercise measures one of those acts and not the other two. Write the reason beside each verdict in the language of what the round measures, and it will still make sense to a reader two years out.
Ask in task terms, and never for a diagnosis
Ask what the round has to change, never what the candidate has. 42 U.S.C. 12112(d)(2)(A) bars an employer from conducting a medical examination or asking a job applicant whether they have a disability, or about its nature or severity, before a conditional offer 2. The question that stays inside that line also gets the better answer: which part of this exercise is the barrier, and what would remove it?
The request has already been made, whatever it sounded like. Under the EEOC's 2022 technical assistance, an applicant who says a medical condition may make a test difficult to take has requested a reasonable accommodation, with no mention of the ADA and without the phrase itself 3.
A request does not have to arrive through a form, either. The sentence counts when it is said to a recruiter, typed into a scheduling thread or mentioned to a vendor's chatbot. A process with a single intake route misses the ones that arrive anywhere else.
29 CFR 1630.2(o)(3) describes what happens next, and its verbs repay a close read. To determine the appropriate reasonable accommodation it may be necessary for the employer to initiate an informal, interactive process with the individual. That process should identify the precise limitations resulting from the disability and the accommodations that could overcome them 5.
The regulation sets no timetable. The candidate's other offer does.
Write down what was asked, what was offered, what was chosen, and what the exercise measured before and after. A rejected candidate who files a charge turns that record into the whole of the response, which is what the position statement has to show about the AI round.
When the tool does touch the round, what do you offer instead?
Offer another route to the same measurement before you offer a refusal. Say the round measures how a candidate frames a problem and defends a call. Framing can be reached out loud, in a longer window, in a different medium, or in a case with the reading burden taken out of it. Decide on a denial only after the alternatives have been looked at.
Four alternatives cover most of what comes up, and none of them needs a new exercise built from scratch:
- More time on the same task, named in the 2022 guidance among accommodations that may be effective 3, which raises its own question about how much time and whether the results stay comparable.
- An oral version, where the candidate talks through the framing and a reviewer writes it down.
- The same case in a different medium: the brief as audio, the answer as a recording.
- A different case that measures the same act with less of the barrier in it.
Cost is the objection that arrives next. The Job Accommodation Network (JAN) surveyed 26,028 employers who had contacted it between January 2019 and December 2024, and drew its report from 5,406 responses 6.
Of the 1,425 employers who gave any cost figure, 61% said the accommodation cost nothing to implement. Another 33% reported a one-time expense, with a median of $300 6.
Every respondent had already called JAN for help with an accommodation, so the sample leans willing. The costs cover workplace accommodations broadly rather than assessments. The figures say accommodations are usually cheap, and say nothing about whether they are simple to arrange.
Swap in an oral version for one person and the exercise changes while the bar stays put. Comparability is the objection to offering an alternative format at all, and it is a real one. No evidence will exist on whether the two versions place people in the same order, because the alternative version will have been run once.
The discipline that holds up is varying the data and never the task or the bar.
The same two questions settle the general case, whether ADA accommodations apply to an AI-based assessment and what to offer. Name the measured act, then ask whether the requested change performs it.
Set the default so the request never has to be made
A round that permits the tool for everyone never produces this request. Put three lines in the invitation: what the exercise measures, which supports are permitted by default, and how to ask for a change. The default version does much of the work an accommodation process does, without asking anybody to disclose anything, and it holds one bar for everyone who gets the brief.
Making the flexible version the default so nobody has to disclose is the design move underneath that, and it costs a paragraph in the invitation.
Then answer the purpose question once, in writing, so it stops arriving one candidate at a time. A round that cannot answer it today has learned that from this request.
The parenthetical in 29 CFR 1630.11 was never about assistive technology.
The words grant one exception, and the exception is a question about purpose: "except where such skills are the factors that the test purports to measure" 1.
Common questions
Can you ask for medical documentation before allowing AI as an accommodation?
Sometimes, and within limits. The EEOC's 2022 technical assistance says an employer may ask for supporting medical documentation where the disability and the need for accommodation are not obvious or already known 3. What stays off limits before a conditional offer is asking whether an applicant has a disability, or about its nature or severity 2. The useful request is functional either way: what the exercise demands that the candidate cannot do in its current form, and what change would remove that. Employment counsel should confirm how this runs in your jurisdiction.
Is a general-purpose model really an accommodation, the way a screen reader is?
The law carries no list of approved devices. A reasonable accommodation is a change that lets a qualified person take part in the application process, and the practical question is whether it works for the limitation, not whether it was sold as accessibility software. So the question is the same one asked of a screen reader: does the requested support perform the specific act the round is measuring? Reading a brief aloud does not. Choosing between two approaches, on an exercise measuring how someone chooses, does.
Does allowing it for one candidate make the round unfair to everyone else?
Only if the change hands the measured act to the candidate who asked. A support that leaves the measurement alone changes the conditions of the exercise and not the bar, which is what an accommodation is meant to do. The genuine risk runs the other way: a ban applied to a whole slate, protecting nothing anybody wrote down, is already grading reading speed and typing next to the judgment it means to score. Naming the measured act in the brief settles the fairness question for every candidate at once.
What if the candidate will not say what the barrier is?
The interactive process runs both ways, and it can stall on either side. 29 CFR 1630.2(o)(3) describes an informal process meant to identify the precise limitations and the accommodations that could overcome them 5, which needs information only the candidate holds. Asking in task terms makes that easier to supply, because it never requires naming a condition. If nothing usable comes back after a written request, record the attempt, the date and the offer that was open, and decide on what you actually have.
Can you deny it because the job itself forbids AI tools?
The job's own restriction is the strongest ground there is, and it has to hold for the work rather than for the round. Where the role genuinely bars the tool (classified work, a client contract barring third-party processing, a data policy no approved model satisfies), an exercise measuring unaided work has its best case for being job-related and consistent with business necessity, the showing 42 U.S.C. 12112(b)(6) asks of the employer 2. Put the restriction in the posting and the brief before anyone applies. A restriction discovered at the moment of the request looks like a rule written to fit one decision.
How much of this has to be written down?
All of it, and it fits in a short note. Record the date of the request, what was asked in task terms, what alternatives were offered, what the candidate chose, and what the exercise measured before and after the change. No diagnosis belongs anywhere in that record. Six months later the note is the only thing that shows the decision turned on the measured act rather than on a hunch about the tool. A charge or an audit would ask for that same note.
References
- 1. 29 CFR 1630.11 - Administration of tests (Regulations to Implement the Equal Employment Provisions of the Americans with Disabilities Act) ✓ govinfo.gov Supports the claim that a test given to an applicant whose disability impairs sensory, manual or speaking skills must reflect what the test purports to measure rather than the impairment, and supports the closing verbatim quotation of the exception clause.
- 2. 42 U.S.C. 12112 - Discrimination ✓ uscode.house.gov Supports two claims: 12112(b)(6) makes it discrimination to use a selection criterion that screens out or tends to screen out an individual with a disability unless job-related and consistent with business necessity, and 12112(d)(2)(A) bars a pre-offer medical examination or inquiry into whether an applicant has a disability or its nature or severity.
- 3. The Americans with Disabilities Act and the Use of Software, Algorithms, and Artificial Intelligence to Assess Job Applicants and Employees ✓ web.archive.org Supports three claims: extended time and an alternative version of a test, including one compatible with a screen reader, are named as accommodations that may be effective; an applicant who says a medical condition may make a test difficult to take has thereby requested a reasonable accommodation without naming the ADA; and the employer may ask for supporting medical documentation where the disability and the need are not obvious or already known.
- 4. Select Issues: Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures Under Title VII of the Civil Rights Act of 1964 (archived capture, 2025-01-25) ✓ web.archive.org Supports the statement that the EEOC's two AI technical assistance documents were taken off eeoc.gov in late January 2025 and read now only through archived captures; this is the archived capture of the companion Title VII document.
- 5. 29 CFR 1630.2 - Definitions ✓ law.cornell.edu Supports the description of the informal, interactive process at 1630.2(o)(3), which should identify the precise limitations resulting from the disability and the potential reasonable accommodations that could overcome them.
- 6. Cost and Benefits of Accommodations (Low Cost, High Impact report) ✓ askjan.org Supports the survey figures: 26,028 employers contacted between January 1, 2019 and December 31, 2024, 5,406 responses, and among the 1,425 who gave cost information, 61% reporting no cost and 33% a one-time expense with a median of $300.
6 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. How Olive sources claims
General guidance, not legal advice. Hiring rules differ by state and country and change often; check anything here against your own counsel before you act on it.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.